Re: [PATCH] mm/khugepaged: Detecting uffd-wp vma more efficiently

From: Axel Rasmussen
Date: Wed Sep 22 2021 - 16:50:31 EST


On Wed, Sep 22, 2021 at 10:52 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
>
> We forbid merging thps for uffd-wp enabled regions, by breaking the khugepaged
> scanning right after we detected a uffd-wp armed pte (either present, or swap).
>
> It works, but it's less efficient, because those ptes only exist for VM_UFFD_WP
> enabled VMAs. Checking against the vma flag would be more efficient, and good
> enough. To be explicit, we could still be able to merge some thps for
> VM_UFFD_WP regions before this patch as long as they have zero uffd-wp armed
> ptes, however that's not a major target for thp collapse anyways.
>
> This mostly reverts commit e1e267c7928fe387e5e1cffeafb0de2d0473663a, but
> instead we do the same check at vma level, so it's not a bugfix.
>
> This also paves the way for file-backed uffd-wp support, as the VM_UFFD_WP flag
> will work for file-backed too.
>
> After this patch, the error for khugepaged for these regions will switch from
> SCAN_PTE_UFFD_WP to SCAN_VMA_CHECK.
>
> Since uffd minor mode should not allow thp as well, do the same thing for minor
> mode to stop early on trying to collapse pages in khugepaged.
>
> Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> Cc: Axel Rasmussen <axelrasmussen@xxxxxxxxxx>
> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: Nadav Amit <nadav.amit@xxxxxxxxx>
> Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> ---
>
> Axel: as I asked in the other thread, please help check whether minor mode will
> work properly with shmem thp enabled. If not, I feel like this patch could be
> part of that effort at last, but it's also possible that I missed something.

Sorry for missing the other thread.

Unfortunately, I think shmem THP *doesn't* really work with minor
faults, and what's worse, just checking the VMA flag isn't enough.

First, let me note the guarantee UFFD minor faults are trying to
provide: for a given mapping, any minor fault (that is, pte_none() but
a page is present in the page cache) must result in a minor userfault
event. Furthermore, the only way the fault may be resolved (i.e., a
PTE installed) is via a UFFDIO_CONTINUE ioctl from userspace.

A typical use case for minor faults is, we have two mappings (i.e.,
two VMAs), both pointing to the same underlying physical memory. It's
typical for both to have MAP_SHARED. It's typical for one of these
mappings to be fully faulted in (i.e., all of its PTEs exist), while
the other one has some missing PTEs. The problem is, khugepaged might
scan *either* of the two mappings. Say it picks the fully-faulted VMA:
even if we set khugepaged_max_ptes_none to zero, it will still go
ahead and collapse these pages - because *this* VMA has no missing
PTEs.

Why is this a problem? When we collapse, we install a PMD, for *all*
VMAs which reference these pages. In other words, we might install
PTEs for the other, minor-fault-registered mapping, and therefore
userfaults will never trigger for some of those regions, even though
userspace never UFFDIO_CONTINUE-ed them.

I *think* the right place to check for this and solve it is in
retract_page_tables(), and I have a patch which does this. I've been
hesitant to send it though, as due to a lack of time and the
complexity involved I haven't been able to write a clear reproducer
program, which my patch clearly fixes. :/

>
> Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> ---
> include/trace/events/huge_memory.h | 1 -
> mm/khugepaged.c | 26 +++-----------------------
> 2 files changed, 3 insertions(+), 24 deletions(-)
>
> diff --git a/include/trace/events/huge_memory.h b/include/trace/events/huge_memory.h
> index 4fdb14a81108..53532f5925c3 100644
> --- a/include/trace/events/huge_memory.h
> +++ b/include/trace/events/huge_memory.h
> @@ -15,7 +15,6 @@
> EM( SCAN_EXCEED_SWAP_PTE, "exceed_swap_pte") \
> EM( SCAN_EXCEED_SHARED_PTE, "exceed_shared_pte") \
> EM( SCAN_PTE_NON_PRESENT, "pte_non_present") \
> - EM( SCAN_PTE_UFFD_WP, "pte_uffd_wp") \
> EM( SCAN_PAGE_RO, "no_writable_page") \
> EM( SCAN_LACK_REFERENCED_PAGE, "lack_referenced_page") \
> EM( SCAN_PAGE_NULL, "page_null") \
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 045cc579f724..3afe66d48db0 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -31,7 +31,6 @@ enum scan_result {
> SCAN_EXCEED_SWAP_PTE,
> SCAN_EXCEED_SHARED_PTE,
> SCAN_PTE_NON_PRESENT,
> - SCAN_PTE_UFFD_WP,
> SCAN_PAGE_RO,
> SCAN_LACK_REFERENCED_PAGE,
> SCAN_PAGE_NULL,
> @@ -467,6 +466,9 @@ static bool hugepage_vma_check(struct vm_area_struct *vma,
> return false;
> if (vma_is_temporary_stack(vma))
> return false;
> + /* Don't allow thp merging for wp/minor enabled uffd regions */
> + if (userfaultfd_wp(vma) || userfaultfd_minor(vma))
> + return false;
> return !(vm_flags & VM_NO_KHUGEPAGED);
> }
>
> @@ -1246,15 +1248,6 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> pte_t pteval = *_pte;
> if (is_swap_pte(pteval)) {
> if (++unmapped <= khugepaged_max_ptes_swap) {
> - /*
> - * Always be strict with uffd-wp
> - * enabled swap entries. Please see
> - * comment below for pte_uffd_wp().
> - */
> - if (pte_swp_uffd_wp(pteval)) {
> - result = SCAN_PTE_UFFD_WP;
> - goto out_unmap;
> - }
> continue;
> } else {
> result = SCAN_EXCEED_SWAP_PTE;
> @@ -1270,19 +1263,6 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> goto out_unmap;
> }
> }
> - if (pte_uffd_wp(pteval)) {
> - /*
> - * Don't collapse the page if any of the small
> - * PTEs are armed with uffd write protection.
> - * Here we can also mark the new huge pmd as
> - * write protected if any of the small ones is
> - * marked but that could bring unknown
> - * userfault messages that falls outside of
> - * the registered range. So, just be simple.
> - */
> - result = SCAN_PTE_UFFD_WP;
> - goto out_unmap;
> - }
> if (pte_write(pteval))
> writable = true;
>
> --
> 2.31.1
>